Integrating individualised and similar group in knowledge tracing

In light of the accelerated growth of online education platforms, Knowledge Tracing (KT), paramount for anticipating learners’ academic performance , assumes an increasingly significant function in real-time content adaptation and forecasting learner outcomes in intelligent education systems . Despite using diverse, complex neural networks to extract features from users’ historical interaction data, researchers often restrict the scope of personalisation and fail to consider the interconnection between individualisation and group similarity. Furthermore, the specific treatment of distinctive characteristics within personalised information is often disregarded in discussions focusing on personalisation. We propose a novel Individualised Group Knowledge Tracing (IGKT) model to address this research gap. The model is designed to focus on learners’ individualised behaviours, employing an attention mechanism to facilitate the processing of significant actions contained within these behaviours. Furthermore, we investigate the problem-skill dimension in conjunction with extracting latent features of learning resources through learners’ study behaviours. In our investigation of group characteristics, we move beyond the conventional aggregation of analogous knowledge states , integrating a more nuanced and detailed set of learning behaviour traits among learners while examining the influence of learning resources on group similarities. The Q-matrix is employed to update learners’ skill mastery levels and learner similarities, thereby integrating personalised and group features in subsequent modules. Furthermore, we conduct a detailed examination of learners’ knowledge state, obtaining a more objective representation of their knowledge state from the perspective of learning resources. We have also designed a forgetting gate incorporating filtered personalised features to achieve an individualised forgetting mechanism. Extensive experiments on three public datasets demonstrate that our model achieves higher prediction accuracy and more precisely captures learners’ knowledge state. Our research findings not only showcase the superiority of our model but also provide valuable insights for future joint studies on personalisation and group characteristics in Knowledge Tracing (KT) models.